The theoretical scaling limit of IBM Power and IBM Z systems is not a single number—it’s the point where adding more CPUs, memory, or I/O no longer increases throughput because of fundamental bottlenecks like serialization, memory bandwidth, interconnect latency, and coherence overhead.

The core idea is:

Scaling is ultimately limited by shared resources and coordination costs, not raw hardware capacity.

1. Two key scaling laws

A. Amdahl's Law (upper bound)

This defines the maximum possible speedup:

• If even 1% of a workload is serial → scaling is capped
• As cores increase → benefit diminishes

👉 Limits strong scaling (fixed workload)

B. Gustafson's Law (practical scaling)

• workload size grows with system size
• allows near-linear scaling for large workloads

👉 More realistic for enterprise systems

2. What “theoretical limit” means in practice

Scaling stops improving when:

Added resources → more contention → no throughput gain

So the limit is where:

• marginal gain ≈ 0
• or latency increases faster than throughput

3. Key bottlenecks that define the limit

A. Memory bandwidth wall

• CPUs demand more data than memory can supply
• becomes dominant at large core counts

👉 very common scaling limit

B. Cache coherence overhead

• maintaining consistency across cores
• increases exponentially with core count

👉 limits SMP scalability

C. Interconnect latency

• communication between sockets/nodes
• affects NUMA and sysplex performance

D. Lock and synchronization contention

• shared data structures
• transaction serialization points

E. I/O subsystem limits

• queueing delays
• storage throughput ceilings

4. IBM Power systems scaling characteristics

A. SMP (Scale-up) design

• large multi-core processors
• shared memory architecture

Theoretical limits:

• hundreds of cores per system
• limited by:
  • cache coherence traffic
  • memory bandwidth

👉 scaling becomes sub-linear beyond high core counts

B. NUMA effects

• remote memory access penalty
• scaling depends on locality

C. SMT interaction

• improves utilization
• but increases contention at scale

5. IBM Z systems scaling characteristics

A. Vertical scaling (single system)

• massive I/O throughput
• specialized processors (zIIP, etc.)
• strong isolation via PR/SM

Limits:

• dispatch contention
• memory bandwidth
• internal fabric limits

B. Horizontal scaling via sysplex

Multiple systems connected:

• shared data via Coupling Facility
• workload distributed across nodes

👉 extends scaling beyond single system

C. Sysplex theoretical limit

Limited by:

• coupling facility latency
• interconnect bandwidth
• lock synchronization overhead

6. Practical scaling numbers (conceptual, not exact)

IBM Power:

• near-linear scaling up to moderate cores
• diminishing returns beyond high core counts
• memory bandwidth becomes dominant

IBM Z:

• very high single-system throughput
• near-linear scaling across sysplex for many workloads
• limited by cross-system coordination

7. Throughput scaling curve

Typical behavior:

1. Linear region
2. Diminishing returns
3. Saturation point
4. Degradation (contention dominates)

8. Why IBM systems scale better than typical systems

A. Advanced cache hierarchy

• large shared caches
• reduces memory pressure

B. High memory bandwidth

• multiple channels
• optimized controllers

C. Hardware-assisted I/O

• offloads CPU
• reduces contention

D. Workload management

• dynamic balancing
• reduces hotspots

E. Partitioning (LPAR)

• isolates workloads
• prevents interference

9. Real theoretical limit (conceptual formula)

Max Throughput ≈ min(
  CPU parallel capacity,
  memory bandwidth,
  I/O bandwidth,
  synchronization throughput,
  interconnect capacity
)

👉 The smallest term defines the limit.

10. Simple mental model

Think of scaling like:

Adding more lanes to a highway helps until all cars must pass through a toll booth—at that point, the toll booth (shared resource) defines the maximum throughput.

11. Key takeaway

The theoretical scaling limit of IBM Power and Z systems is determined by:

• Amdahl’s Law (serial fraction)
• memory bandwidth saturation
• cache coherence overhead
• interconnect latency
• synchronization and lock contention
• I/O throughput limits

👉 Even in highly advanced systems, shared resource contention—not CPU count—defines the ultimate scaling boundary